ObjectivePredicting soil nutrients by visible-near infrared (vis-NIR) and mid-infrared (MIR) spectroscopy has the advantages of being fast, cost-effective and environmental friendly. Soil spectra contain abundant information of soil properties, and can be combined with machine learning methods to effectively and accurately predict soil nutrients, which can provide support and guidance for timely fertilization management. The objective of this study was to compare the predictive ability of vis-NIR (350-2500 nm) and MIR spectroscopy (4000-650 cm^-1) for predicting both the total and available contents of soil nitrogen (N), phosphorus (P) and potassium (K), in order to construct an optimal model for estimation of different nutrient contents.MethodIn this study, 500 samples were collected from the surface layers (0-20 cm) of the dryland in Guizhou Province for determination of soil N, P and K contents and spectral analysis. The vis-NIR spectra were measured by Cary 5000 and the MIR spectra by Thermo Scientifit Nicolet iS50. Soil spectra were pre-processed by Savitzky-Golay (SG) smoothing for denoising and standard normal variate (SNV) transformation for baseline correction. Partial least squares regression (PLSR) and support vector machine (SVM) were used to predict the contents of total nitrogen (TN), total phosphorus (TP), total potassium (TK), alkali-hydrolyzable nitrogen (AN), available phosphorus (AP) and available potassium (AK).ResultThe results showed that: (1) Whether using the vis-NIR spectroscopy or the MIR spectroscopy, the prediction accuracy of PLSR model was better than that of SVM model. (2) The accuracy of MIR spectroscopy for prediction of TN, TK and AN was significantly higher than that of vis-NIR spectroscopy. Vis-NIR and MIR spectroscopy could reliably predict TN and TK(ratio of performance to interquartile distance (RPIQ) > 2.10), while MIR spectroscopy could predict AN with moderate accuracy (RPIQ = 1.87). However, both types of spectra had poor ability to predict TP, AP and AK (RPIQ < 1.34). (3) When the variable in the projection (VIP) score was > 1.5, there were more important bands selected by PLSR models in the MIR region than the vis-NIR region. The important bands selected for estimation of TN were mainly concentrated near 1910 and 2207 nm in the vis-NIR region, and centered around 1 120, 1 000, 960, 910, 770, and 668 cm^-1 in the MIR region. The important bands of TK were mainly distributed around 540, 2 176, 2 225, and 2 268 nm in the vis-NIR region, and around 1 040, 960, 910, 776, 720, and 668 cm^-1 in the MIR region.ConclusionTherefore, MIR spectroscopy combined with PLSR model proved to be promising for accurate prediction of soil nutrients, especially for the estimation of TN and TK, and can provide technical support for guiding timely fertilization.